Enhance dataset card: Add description, links, and task category

This PR significantly enhances the GeoBench dataset card by adding:
- A clear descriptive overview of the dataset's purpose and content.
- A link to the associated paper: https://huggingface.co/papers/2511.15705
- A link to the project page: https://ekonwang.github.io/geo-vista/
- A link to the code repository: https://github.com/ekonwang/GeoVista
- The `task_categories: ['image-text-to-text']` metadata, improving the dataset's discoverability and classification on the Hugging Face Hub.

These additions provide essential context for users and researchers interested in geolocalization and agentic visual reasoning.

README.md

@@ -16,4 +16,15 @@ configs:
   data_files:
   - split: train
     path: data/train-*
+task_categories:
+- image-text-to-text
 ---
+
+# GeoBench Dataset for GeoVista
+
+The **GeoBench** dataset is a benchmark introduced in the paper [GeoVista: Web-Augmented Agentic Visual Reasoning for Geolocalization](https://huggingface.co/papers/2511.15705).
+
+GeoBench is curated to rigorously evaluate the geolocalization ability of agentic models. It comprises photos and panoramas from around the world, as well as a subset of satellite images of different cities. This dataset addresses the demand for high-resolution imagery and the complexities of localization for advanced agentic reasoning, necessitating nuanced visual grounding and web search to validate hypotheses during the reasoning process.
+
+- **Project Page:** https://ekonwang.github.io/geo-vista/
+- **Code:** https://github.com/ekonwang/GeoVista